CLOSE NEVER - Avoid pseudo-close for active datasets and in-memory objects, to avoid sync I/O due to XI

PROBLEM:

Our shop has an active 10-way data-sharing environment and we have been fighting for some time to get our group bufferpool Sync/XI hit ratios up above 90%. We appear to be fighting "pseudo-close" in this endeavor.

According to our conversations with IBM, during pseudo-close, Db2 cross-invalidates large swaths of pages from local bufferpools...they do this because serialization during externalization can cause concurrency issues. However, these invalid pages lead to large numbers of Sync-for-XI misses (and thus sync I/Os) for data sharing. Two examples:

SCENARIOS:

1) We have a large index bufferpool, about 8GB in each subsystem, with a 20GB group bufferpool. Based on update rates, the GBP should have plenty of pages to cache XI'd pages for long periods of time...yet our Sync XI hit ratios fall down to 70% or below. In discussion with IBM, we were advised to set PCLOSET higher to avoid pseudo-close. We set it to 60 minutes, and this appeared to help for a while, but now Sync/XI hit ratios are drifting down again. Setting PCLOSET much higher could compromise recovery for objects, and many objects in the environment do NOT require this level of performance.

2) We are operating an "in-memory" (PGSTEAL NONE) bufferpool 20GB in size. The storage is MORE than enough to cache all objects assigned to it...yet after a while we start seeing sync I/O "trickles" and "bursts". After discussion with IBM, we found that pseudo-close on an infrequent-updating member was forcing XIs on all members, resulting in sync I/O to retrieve pages from disk.

THE PROPOSED SOLUTION:

Offer a "CLOSE NEVER" setting for tablespaces and indexes, in addition to CLOSE NO and CLOSE YES. This would instruct Db2 not to close these pagesets for write (not pseudo-close) unless they were either quiesced/started RO or DB2 was shut down.

CLOSE NEVER objects would never pseudo-close, so pages would stay resident in the GBP until "stolen" and we we would not need to incur sync I/O due to pseudo-close cross-invalidations. This should allow for higher Sync XI hit ratios for heavily used objects.

This would be especially useful for "in-memory" data in PGSTEAL(NONE) bufferpools. Eliminating the XIs due to pseudoclose/RO conversion would eliminate much of the I/O and allow the GBPs to cover those pages that were XI'd due to normal write activity, keeping in-memory data actually "in-memory".

CONSIDERATIONS:

Users would need to be notified in documentation for CLOSE NEVER:

* Keep a close eye on DSMAX, as these datasets would reduce the number of DSCBs available for other data.

* Make sure to have a good backup/recovery strategy, as SYSLGRNX intervals may not close. (alternately, give users the ability to close a SYSLGRNX interval manually for CLOSE NEVER data without XI-ing the GBPs, but that could be tricky)